Hello Danijela and congratulation on completing your masters degree. In your master thesis, you test how good are forecasts of the Universal Thermal Climate Index (UTCI), an index describing the physiological response of a human body to an actual state of the environment (a combination of meteorological parameters such as temperature, humidity, wind, and radiation) in terms of temperature, for Slovenia. Your thesis spans multiple fields of meteorology, including climatology, biometeorology, and numerical weather prediction. The title of your thesis is »Verification of Operational Universal Thermal Climate Index Forecasts for Slovenia.« Your advisor was assoc. prof. dr. Gregor Skok (FMF) and your co-advisor was dr. Jana Banko (Slovenian Environment Agency – ARSO).
Please present the background and research question of your thesis in a short and simple manner.
Hello, first of all, thank you for your sincere congratulations. As you mentioned, in my master’s thesis I tested how good the UTCI index forecast is and tried to improve it with post-processing methods- neural networks and linear regression. Generally, post-processing is a process, in which the raw output from a numerical weather model is analyzed, filtered, and adjusted to provide a more precise representation of the weather conditions.
UTCI is one of the thermal indices based on a heat balance model, considering the meteorological factors and the influence of clothing. Every day, we witness how weather can affect a person’s well-being and health, especially extreme events, which can also be dangerous to human life. Recently, extreme weather events are happening more often and with greater intensity. In order to express unified environmental influences on humans throughout the entire range of thermal conditions, for all climates and seasons, UTCI was developed.
This masters thesis is the result of collaboration with the Slovenian Environment Agency (ARSO). UTCI was recently implemented in the ALADIN model, a numerical weather prediction model used at ARSO, and operationally published as part of the bioweather forecast, daily available to users on the ARSO website. They provided us with archived operational outputs of ALADIN’s UTCI forecasts for the period 2013-2018 and for the year 2022. We verified it based on the measured meteorological data, where the UTCI values were calculated with BioKlima model, software for the calculation of different biometeorological and thermophysiological indices.
The measured meteorological data came from 42 stations in Slovenia, where the radiation measurements were performed. For Slovenia, these tests have not yet been carried out. Additionally, we also performed an overview of UTCI climatology, to get a general picture of Slovenian thermal conditions.
The secondary goal of the master thesis was an attempt to improve the operational UTCI forecast with the help of two relatively simple methods, linear regression, and dense sequential neural network. The aim was potential implementation in the operational process at ARSO in the future if one of the methods significantly improves the quality of the forecast. We had six different setups of neural networks, which were the same, except for the number of input parameters. The evaluation of neural networks was done by comparing their mean absolute error (MAE) with respect to the observations for the test cases.
What are the main results and conclusions of your thesis?
From our research, we can conclude that the ALADIN model tends to overestimate UTCI. On average, the daily mean error is 2.57°C, while the mean absolute error is 5.023°C. The mean error reaches its maximum around 07:00 h when it almost reaches 8°C. After that, during the afternoon it decreases below 0°C, to -2°C. The average mean absolute error also has a peak around 07:00h, when it rises above 8°C. The hourly average mean error (ME) in the period of the day when we expect maximal UTCI is in the range of the average ME. There is a small percentage of cases with very large errors. We found that the possible sources of large errors may be: incorrectly forecasted value of wind speed or temperature, incorrectly forecasted cloud cover, humidity, and surface radiation balance, a large difference between the model and actual altitude of the stations, specific meteorological conditions at locations of meteorological stations, possible errors in measurements, approximation of sun elevation determination in BioKlima, which could lead to the timing problems, and unknown approximations in BioKlima.
During the second part of my master’s thesis, we made successful improvements to ALADIN’s forecasts of UTCI using both post-processing methods. The neural network method outperformed linear regression, but not quite to the extent we had expected. The hourly course of ME was also flatter and better, but the average ME for the neural network method was slightly higher than the linear regression method in all examples. We were able to achieve the best set of input data resulting in an MAE of 3.0271 for the neural network method and an MAE of 3.4715 for linear regression, as compared to ALADIN’s MAE of 5.023.
Although we have observed many problems and limitations, this study can serve as a good start for further verification, application, and improvement of ALADIN’s UTCI forecasts. We have seen that UTCI is one of the more demanding indices that involves many processes, and that is why its application is wide and significant. Its verification is difficult due to a lack of instruments for direct measurements of UTCI or of mean radiant temperature. Besides that, it would be useful to have open-source code in the popular system, independent of the operating system. Based on the results of the second part of this thesis, the neural network method of post-processing could be potentially implemented in the UTCI forecast at ARSO.
Thanks for the replies and good luck
Link to PDF of the master thesis
Interview by: assist. dr. Katarina Kosovelj
